Kris Szalay
kristofszalay.com
Kris Szalay
@kristofszalay.com
CTO @ Turbine
We did some tests to see what's possible with #virtualcells today with all the headlines. Turns out what's available for #training is what really matters, not the assay. Like #AGI to #AI, what we have today is narrow #virtual #cells. But there is a path to generality.
blog.turbine.ai/p/what-can-v...
What can Virtual Cells do for you today?
and what they can't - because the bottleneck is not computing power
blog.turbine.ai
November 11, 2025 at 8:12 PM
There such a big disconnect between how #pharma insiders view the value of #AI and what's in the news. It's because retrospective #clinical #data prediction is very different than for a new #drug.
blog.turbine.ai/p/the-patient-puzzle
Thanks to #MSK and #Champions for supporting us in this effort!
The Patient Puzzle
on why solving patients need more than patients
blog.turbine.ai
October 16, 2025 at 9:33 PM
The reason doing #ML in #biology is much harder than doing ML on text is that "ground truth" is very different.
In the post below I'll show you the first 5 things we check when curating one single data point to teach the right idea to the #AI. #compbio #virtualcell

blog.turbine.ai/p/ic50-is-a-...
IC50 is a deep rabbit hole
On harmonizing data, because ground truth in biology is unlike text
blog.turbine.ai
June 30, 2025 at 2:35 PM
Why can #AI work in #biology?

It's not about hidden shortcuts to do raw physics. I think it's because we are learning how evolution was shaping us over all this time to survive.
blog.turbine.ai/p/why-can-we...
So maybe we should move on from individual proteins to learning relations of processes.
Why can we predict biology?
and what are we really predicting?
blog.turbine.ai
May 22, 2025 at 7:27 PM
Reposted by Kris Szalay
Single-cell foundation models, trained on large-scale scRNA-seq datasets, are increasingly used for post-perturbation RNA-seq prediction.

But how do they actually perform?

Our new paper from @turbine-ai.bsky.social is now out in BMC Genomics. bmcgenomics.biomedcentral.com/articles/10....

1/5
April 25, 2025 at 9:49 PM
Turns out the way we usually pre-train foundational cell models adds very little information to the system - definitely not enough to make drug effect predictions work. Not what I've expected.
#virtualcells #foundationalmodels #compbio

blog.turbine.ai/p/pretrainin...
Pretraining virtual cells is useless
the way we do it now
blog.turbine.ai
April 16, 2025 at 5:03 PM
Reposted by Kris Szalay
The 'Google Co-Scientist' claims to be 'Accelerating scientific breakthroughs with an AI co-scientist'

But does it?

A closer look at its scientific validation as currently reported

Please share and let me know if you have any comments, cheers! Andreas

www.drugdiscovery.net/2025/02/20/t...
February 20, 2025 at 6:58 AM
also check out blog.turbine.ai/archive :)
Archive - Turbine AI
Full archive of all the posts from Turbine AI.
blog.turbine.ai
February 5, 2025 at 9:49 PM
printf("hello world!");
February 5, 2025 at 9:45 PM